LibrettOS: A Dynamically Adaptable Multiserver-Library OS

Page created by Gerald Chapman
 
CONTINUE READING
LibrettOS: A Dynamically Adaptable Multiserver-Library OS∗
                                                                              Ruslan Nikolaev, Mincheol Sung, Binoy Ravindran
                                                            Bradley Department of Electrical and Computer Engineering, Virginia Tech
                                                                              {rnikola, mincheol, binoy}@vt.edu

                                         Abstract                                                                     network components fail or need to be upgraded. Finally, to
arXiv:2002.08928v1 [cs.OS] 20 Feb 2020

                                                                                                                      efficiently use hardware resources, applications can dynam-
                                         We present LibrettOS, an OS design that fuses two                            ically switch between the indirect and direct modes based
                                         paradigms to simultaneously address issues of isolation,                     on their I/O load at run-time. We evaluate LibrettOS with
                                         performance, compatibility, failure recoverability, and run-                 10GbE and NVMe using Nginx, NFS, memcached, Redis, and
                                         time upgrades. LibrettOS acts as a microkernel OS that runs                  other applications. LibrettOS’s performance typically ex-
                                         servers in an isolated manner. LibrettOS can also act as a                   ceeds that of NetBSD, especially when using direct access.
                                         library OS when, for better performance, selected applica-
                                                                                                                      Keywords: operating system, microkernel, multiserver, net-
                                         tions are granted exclusive access to virtual hardware re-
                                                                                                                      work server, virtualization, Xen, IOMMU, SR-IOV, isolation
                                         sources such as storage and networking. Furthermore, ap-
                                         plications can switch between the two OS modes with no
                                         interruption at run-time. LibrettOS has a uniquely distin-
                                         guishing advantage in that, the two paradigms seamlessly
                                                                                                                      1     Introduction
                                         coexist in the same OS, enabling users to simultaneously ex-
                                                                                                                      Core components and drivers of a general purpose mono-
                                         ploit their respective strengths (i.e., greater isolation, high
                                                                                                                      lithic operating system (OS) such as Linux or NetBSD typ-
                                         performance). Systems code, such as device drivers, net-
                                                                                                                      ically run in privileged mode. However, this design is of-
                                         work stacks, and file systems remain identical in the two
                                                                                                                      ten inadequate for modern systems [13, 30, 43, 50, 62]. On
                                         modes, enabling dynamic mode switching and reducing de-
                                                                                                                      the one hand, a diverse and ever growing kernel ecosys-
                                         velopment and maintenance costs.
                                                                                                                      tem requires better isolation of individual drivers and other
                                            To illustrate these design principles, we implemented a                   system components to localize security threats due to the
                                         prototype of LibrettOS using rump kernels, allowing us to                    increasingly large attack surface of OS kernel code. Bet-
                                         reuse existent, hardened NetBSD device drivers and a large                   ter isolation also helps with tolerating component failures
                                         ecosystem of POSIX/BSD-compatible applications. We use                       and thereby increases reliability. Microkernels achieve this
                                         hardware (VM) virtualization to strongly isolate different                   goal, specifically in multiserver OS designs [13, 30, 33].1 On
                                         rump kernel instances from each other. Because the orig-                     the other hand, to achieve better device throughput and
                                         inal rumprun unikernel targeted a much simpler model for                     resource utilization, some applications need to bypass the
                                         uniprocessor systems, we redesigned it to support multi-                     system call and other layers so that they can obtain exclu-
                                         core systems. Unlike kernel-bypass libraries such as DPDK,                   sive access to device resources such as network adapter’s
                                         applications need not be modified to benefit from direct                     (NIC) Tx/Rx queues. This is particularly useful in recent
                                         hardware access. LibrettOS also supports indirect access                     hardware with SR-IOV support [63], which can create vir-
                                         through a network server that we have developed. Instances                   tual PCIe functions: NICs or NVMe storage partitions. Li-
                                         of the TCP/IP stack always run directly inside the address                   brary OSes and kernel-bypass libraries [77, 84] achieve this
                                         space of applications. Unlike the original rumprun or mono-                  goal. Multiserver-inspired designs, too, can outperform tra-
                                         lithic OSes, applications remain uninterrupted even when                     ditional OSes on recent hardware [54, 56].
                                             ∗ ©2020 Copyright held by the owner/author(s). Publication rights           Microkernels, though initially slow in adoption, have gai-
                                         licensed to ACM. This is the author’s version of the work. It is             ned more traction in recent years. Google’s Fuchsia OS [24]
                                         posted here for your personal use. Not for redistribution. The defini-       uses the Zircon microkernel. Intel Management Engine [37]
                                         tive Version of Record was published in Proceedings of the 16th              uses MINIX 3 [30] since 2015. A multiserver-like network-
                                         ACM SIGPLAN/SIGOPS International Conference on Virtual Execu-
                                         tion Environments (VEE ’20), March 17, 2020, Lausanne, Switzerland
                                                                                                                      ing design was recently revisited by Google [56] to improve
                                         http://dx.doi.org/10.1145/3381052.3381316.                                   performance and upgradability when using their private

                                         The U.S. Government is authorized to reproduce and distribute reprints          1 As microkernels are defined broadly in the literature, we clarify that

                                         for Governmental purposes notwithstanding any copyright annotation           we consider multiserver OSes as those implementing servers to isolate core
                                         thereon.                                                                     OS components, e.g., MINIX 3 [30].

                                                                                                                  1
devices restrict the number of SR-IOV interfaces – e.g., the
                   storage)
                                                                    JDK
             (NVMestorage)
                                     python
                                                                              Intel 82599 adapter [38] supports up to 16 virtual NICs, 4
                                      NFS                           ntp       Tx/Rx queues each; for other adapters, this number can be
                                                     Network
                                     rsyslog          Server        ssh       even smaller. Thus, it is important to manage available
         bus(NVMe

                                                                              hardware I/O resources efficiently. Since I/O load is usu-
                                DB       key-value              HTTP
                                                               (Nginx)        ally non-constant and changes for each application based
                              (mySQL)     (Redis)
    PCIebus

                                                                              on external circumstances (e.g., the number of clients con-
                                                                              nected to an HTTP server during peak and normal hours),
    PCIe

                                            PCIe
                                            PCIebus
                                                 bus(NIC)
                                                     (NIC)                    conservative resource management is often desired: use
                                                                              network server(s) until I/O load increases substantially, at
                              Figure 1. Server ecosystem example.             which point, migrate to the library OS mode (for direct ac-
                                                                              cess) at run-time with no interruption. This is especially
                                                                              useful for recent bare metal cloud systems – e.g., when one
(non-TCP) messaging protocol. However, general-purpose                        Amazon EC2 bare metal instance is shared by several users.
application and device driver support for microkernels is                        In this paper, we present a new OS design – LibrettOS
limited, especially for high-end hardware such as 10GbE+.                     – that not only reconciles the library and multiserver OS
                                                                              paradigms while retaining their individual benefits (of bet-
   Kernel-bypass techniques, e.g., DPDK [84] and                              ter isolation, failure recoverability, and performance), but
SPDK [77], are also increasingly gaining traction, as                         also overcomes their downsides (of driver and application
they eliminate OS kernels from the critical data path,                        incompatibility). Moreover, LibrettOS enables applications
thereby improving performance. Unfortunately, these                           to switch between these two paradigms at run-time. While
techniques lack standardized high-level APIs and require                      high performance can be obtained with specialized APIs,
massive engineering effort to use, especially to adapt to                     which can also be adopted in LibrettOS, they incur high en-
existing applications [89]. Additionally, driver support in                   gineering effort. In contrast, with LibrettOS, existing appli-
kernel-bypass libraries such as DPDK [84] is great only                       cations can already benefit from more direct access to hard-
for high-end NICs from certain vendors. Re-implementing                       ware while still using POSIX.
drivers and high-level OS stacks from scratch in user space                      We present a realization of the LibrettOS design through
involves significant development effort.                                      a prototype implementation. Our prototype leverages rump
   Oftentimes, it is overlooked that “no one size fits all.” In               kernels [42] and reuses a fairly large NetBSD driver collec-
other words, no single OS model is ideal for all use cases.                   tion. Since the user space ecosystem is also inherited from
Depending upon the application, security or reliability re-                   NetBSD, the prototype retains excellent compatibility with
quirements, it is desirable to employ multiple OS paradigms                   existing POSIX and BSD applications as well. Moreover, in
in the same OS. In addition, applications may need to switch                  the two modes of operation (i.e., library OS mode and multi-
between different OS modes based on their I/O loads. In                       server OS mode), we use an identical set of drivers and soft-
Figure 1, we illustrate an ecosystem of a web-driven server                   ware. In our prototype, we focus only on networking and
running on the same physical or virtual host. The server                      storage. However, the LibrettOS design principles are more
uses tools for logging (rsyslog), clock synchronization (ntp),                general, as rump kernels can potentially support other sub-
NFS shares, and SSH for remote access. The server also runs                   systems – e.g., NetBSD’s sound drivers [18] can be reused.
python and Java applications. None of these applications                      The prototype builds on rumprun instances, which execute
are performance-critical, but due to the complexity of the                    rump kernels atop a hypervisor. Since the original rumprun
network and other stacks, it is important to recover from                     did not support multiple cores, we redesigned it to support
temporary failures or bugs without rebooting, which is im-                    symmetric multiprocessing (SMP) systems. We also added
possible in monolithic OSes. One way to meet this goal is                     10GbE and NVMe drivers, and made other improvements to
to have a network server as in the multiserver paradigm,                      rumprun. As we show in Section 5, the prototype outper-
which runs system components in separate user processes                       forms the original rumprun and NetBSD, especially when
for better isolation and failure recoverability. This approach                employing direct hardware access. In some tests, the pro-
is also convenient when network components need to be                         totype also outperforms Linux, which is often better opti-
upgraded and restarted at run-time [56] by triggering an ar-                  mized for performance than NetBSD.
tificial fault.                                                                  The paper’s research contribution is the proposed OS
   Core applications such as an HTTP server, database, and                    design and its prototype. Specifically, LibrettOS is the first
key-value store are more I/O performance-critical. When                       OS design that simultaneously supports the multiserver and
having a NIC and NVMe storage with SR-IOV support (or                         library OS paradigms; it is also the first design that can dy-
multiple devices), selected applications can access hardware                  namically switch between these two OS modes at run-time
resources directly as in library OSes or kernel-bypass li-                    with no interruption to applications. Our network server,
braries – i.e., by bypassing the system call layer or inter-                  designed for the multiserver mode, accommodates both
process communication (IPC). Due to finite resources, PCIe                    paradigms by using fast L2 frame forwarding from appli-

                                                                          2
cations. Finally, LibrettOS’s design requires only one set of          2.2   NetBSD, rump kernels, and rumprun
device drivers for both modes and mostly retains POSIX/BSD
                                                                       Rump kernels are a concept introduced and popularized by
compatibility.
                                                                       Antti Kantee [42] and the larger NetBSD [59] community.
   Our work also confirms an existent hypothesis that ker-             NetBSD is a well known monolithic OS; along with FreeBSD
nel bypass is faster than a monolithic OS approach. How-               and OpenBSD, it is one of the most popular BSD systems
ever, unlike prior works [7, 56, 64], we show this without             in use today. NetBSD provides a fairly large collection of
resorting to optimized (non-POSIX) APIs or specialized de-             device drivers, and its latest 8.0 version supports state-of-
vice drivers.                                                          the-art 10GbE networking, NVMe storage, and XHCI/USB
                                                                       3.0. Unlike in other OSes, NetBSD’s code was largely re-
                                                                       designed to factor out its device drivers and core compo-
2     Background                                                       nents into anykernel components. As shown in Figure 2, a
                                                                       special rump kernel glue layer enables execution of anyker-
For greater clarity and completeness, in this section, we dis-         nels outside of the monolithic NetBSD kernel. For example,
cuss various OS designs, microkernels, and hypervisors. We             using POSIX threads (pthreads), rump kernels can run anyk-
also provide background information on rump kernels and                ernels in user space on top of existing OSes. In fact, NetBSD
their practical use for hypervisors and bare metal machines.           already uses rump kernels to facilitate its own device driver
                                                                       development and testing process, which is done much easier
                                                                       in user space.
2.1    Multiserver and Library OSes                                                                                rump kernel
                                                                                                                   anykernel

                                                                                                calls
Traditionally, monolithic OSes run all critical components

                                                                              Program
in a privileged CPU mode within a single kernel address                                                    USB       VFS         TCP/IP

                                                                              Program

                                                                                         libc
                                                                                        libc
space that is isolated from user applications running in
                                                                                                           XHCI     NVMe          NIC

                                                                                                function
an unprivileged mode. The isolation of systems software
components such as device drivers in separate address
spaces is the fundamental principle behind the microker-                                                      rump
                                                                                                               rumpglue
                                                                                                                    gluecode
                                                                                                                         code(**)
                                                                                                                              (**)
nel model [1, 27, 30, 52], which provides stronger security
and reliability [22, 31, 46] guarantees to applications com-                                                      rumprun
                                                                                                                   rumprun(*)
                                                                                                                           (*)
pared to the classical monolithic kernel designs. To support
existing applications and drivers, microkernel systems ei-             Figure 2. Rump software stack, (*) indicates components
ther implement emulation layers [26] or run a multiserver              substantially and (**) partially modified in LibrettOS.
OS [13, 20, 29, 30, 33] on top of the microkernel. In the multi-
server design, servers are typically created for specific parts           Rump kernels may serve as a foundation for new OSes, as
of the system, such as device drivers and network and stor-            in the case of the rumprun unikernel. As shown in Figure 2,
age stacks. Applications communicate with the servers us-              systems built from rump kernels and rumprun are, effec-
ing IPC to access hardware resources.                                  tively, library OSes. In contrast to NetBSD, rumprun-based
   The system call separation layer between user applica-              systems substitute system calls with ordinary function calls
tions and critical components in monolithic OSes is some-              from libc. The original rumprun, unfortunately, lacks SMP
what arbitrary, chosen for historical reasons such as rough            support.
correspondence to the POSIX standard. However, more ker-                  Rump kernels were also used to build servers as in the
nel components can be moved into the application itself,               rump file system server [41]. However, prior solutions sim-
bringing it closer to the hardware. This concept was ini-              ply rely on user-mode pthreads, targeting monolithic OSes.
tially proposed by Tom Anderson [3], later implemented                 The design of such servers is very different from a more
in the exokernel model [15], and named the “library OS.”               low-level, microkernel-based architecture proposed in this
This model is often advocated for performance reasons and              paper.
OS flexibility [6]. Subsequent works pointed out the secu-
rity benefits of the library OS model [66, 85, 86] due to the
                                                                       2.3   Linux-based Library OSes
strong isolation provided by the reduced interactions be-
tween the application and the privileged layer. Specialized            We also considered Linux Kernel Library (LKL) [67], which
library OSes such as unikernels [10, 43, 45, 50, 55, 90] do            is based on the Linux source. However, LKL is not as flexible
not isolate OS components from applications and are de-                as rump kernels yet. Additionally, rump kernels have been
signed to run just a single application in virtualized envi-           upstreamed into the official NetBSD repository, whereas
ronments typical of cloud infrastructures. As unikernels are           LKL is still an unofficial fork of the Linux source branch.
application-oriented, they may provide a significant reduc-            rump kernels also support far more applications at the mo-
tion in software stacks compared to full-fledged OSes.                 ment.

                                                                   3
System               API                    Generality         TCP                 Driver     Paradigms      Dynamic      Direct Failure
                                                                 Stack               Base                      Switch       Access Recovery
  DPDK [84]            Low-level              Network            3rd party           Medium     N/A            No           Yes       No
  SPDK [77]            Low-level              Storage            N/A                 NVMe       N/A            No           Yes       No
  IX [7]               Non-standard           Network            App                 DPDK       Library OS     No           Almost    No
  Arrakis [64]         POSIX/Custom           Network            App                 Limited    Library OS     No           Yes       No
  Snap [56]            Non-standard           Network            unavailable         Limited    Multiserver    No           No        Yes
  VirtuOS [62]         POSIX/Linux            Universal          Server              Large      Multiserver    No           No        Yes
  MINIX 3 [30]         POSIX                  Universal          Server              Limited    Multiserver    No           No        Yes
  HelenOS [13]         POSIX                  Universal          Servers             Limited    Multiserver    No           No        Yes
  Linux                POSIX/Linux            Universal          Kernel              Large      Monolithic     No           No        No
  NetBSD [59]          POSIX/BSD              Universal          Kernel              Large      Monolithic     No           No        No
  LibrettOS            POSIX/BSD              Universal          App                 Large      Multiserver,   Yes          Yes       Yes
                                                                                     (NetBSD)   Library OS

    Table 1. Comparison of LibrettOS with libraries and frameworks as well as monolithic, multiserver, and library OSes.

   Unikernel Linux (UKL) [69] uses the unikernel model for                            where rumprun instances execute as ordinary microkernel
Linux by removing CPU mode transitions. UKL is in an                                  processes. The seL4 version of rumprun still lacks SMP sup-
early stage of development; its stated goals are similar to                           port, as it only changes its platform abstraction layer.
that of rumprun. However, it is unclear whether it will even-
tually match rump kernels’ flexibility.
   Aside from technical issues, due to incompatibility,                               2.5     PCI Passthrough and Input-Output
Linux’s non-standard GPLv2-only license may create le-                                        Memory Management Unit (IOMMU)
gal issues2 when linking proprietary3 or even GPLv3+4                                 VMs can get direct access to physical devices using their
open-source applications. Since licensing terms cannot be                             hypervisor’s PCI passthrough capabilities. IOMMU [2, 39]
adapted easily without explicit permission from all numer-                            hardware support makes this process safer by remapping in-
ous Linux contributors, this can create serious hurdles in                            terrupts and I/O addresses used by devices. When IOMMU
the wide adoption of Linux-based library OS systems. In                               is enabled, faulty devices are unable to inject unassigned
contrast, rump kernels and NetBSD use a permissive 2-BSD                              interrupts or perform out-of-bound DMA operations. Sim-
license.                                                                              ilarly, IOMMU prevents VMs from specifying out-of-bound
                                                                                      addresses, which is especially important when VM appli-
2.4     Hypervisors and Microkernels                                                  cations are given direct hardware access. We use both PCI
                                                                                      passthrough and IOMMU in the LibrettOS prototype to ac-
Hypervisors are typically used to isolate entire guest OSes                           cess physical devices.
in separate Virtual Machines (VMs). In contrast, microker-
nels are typically used to reduce the Trusted Computing Base
(TCB) of a single OS and provide component isolation us-                              2.6     Single-Root I/O Virtualization (SR-IOV)
ing process address space separation. Although hypervisors
                                                                                      To facilitate device sharing across isolated entities such as
and microkernels have different design goals, they are inter-
                                                                                      VMs or processes, self-virtualizing hardware [68] can be
related and sometimes compared to each other [26, 28].
                                                                                      used. SR-IOV [63] is an industry standard that allows split-
Moreover, hypervisors can be used for driver isolation in
                                                                                      ting one physical PCIe device into logically separate vir-
VMs [17], whereas microkernels can be used as virtual ma-
                                                                                      tual functions (VFs). Typically, a host OS uses the device’s
chine monitors [16].
                                                                                      physical function (PF) to create and delete VFs, but the PF
   The original uniprocessor rumprun executes as a par-
                                                                                      need not be involved in the data path of the VF users. SR-
avirtualized (PV) [88] guest OS instance on top of the Xen
                                                                                      IOV is already widely used for high-end network adapters,
hypervisor [4] or KVM [44]. Recently, rumprun was also
                                                                                      where VFs represent logically separate devices with their
ported [14] to seL4 [46], a formally verified microkernel,
                                                                                      own MAC/IP addresses and Tx/Rx hardware buffers. SR-
   2 This is just an opinion of this paper’s authors and must not be inter-           IOV is also an emerging technology for NVMe storage,
preted as an authoritative legal advice or guidance.                                  where VFs represent storage partitions.
   3 Due to imprecise language, Linux’s syscall exception falls into a gray

legal area when the system call layer is no longer present or clearly defined.
                                                                                         VFs are currently used by VMs for better network perfor-
   4 Without the syscall exception, GPLv3+ is incompatible with GPLv2                 mance, as each VM gets its own slice of the physical device.
(GPLv2+ only) [83]. Moreover, most user-space code has moved to GPLv3+.               Alternatively, VFs (or even the PF) can be used by the DPDK

                                                                                 4
library [84], which implements lower layers of the network                 Application 1, libOS              Application 2                               Network
                                                                             (direct access)                (uses servers)                                Server
stack (L2/L3) for certain NIC adapters. High-throughput
                                                                                 M:N
network applications use DPDK for direct hardware access.                    scheduler                                                                      M:N
                                                                                    libc                              libc                               scheduler

2.7    POSIX and Compatibility                                                FS           TCP/IP                  TCP/IP
                                                                                                                                         IPC
Although high-throughput network applications can bene-                      NVMe            NIC                      VIF                               NIC
fit from custom APIs [7], POSIX compatibility remains crit-                        rumprun                         rumprun                          rumprun
ical for many legacy applications. Moreover, modern server
applications require asynchronous I/O beyond POSIX, such                                                    μ-kernel or hypervisorkernel or hypervisor
                                                                                scheduler                                          (Xen, seL4, etc)
as Linux’s epoll_wait(2) or BSD’s kqueue(2) equivalent.
                                                                                                    PCI passthrough (with IOMMU)
Typically, library OSes implement their own APIs [7, 64].                                  PCIe bus (PF or VF through SR-kernel or hypervisorIOV)
Some approaches [34, 85] support POSIX but still run on top
of a monolithic OS with a large TCB. LibrettOS provides full
                                                                                    Figure 3. LibrettOS’s architecture.
POSIX and BSD compatibility (except fork(2) and pipe(2) as
discussed in Section 4.2), while avoiding a large TCB in the
underlying kernel.
                                                                      Since an OS typically runs several applications, hardware
                                                                      resources need to be shared. In this model, devices are
2.8    Summary                                                        shared using hardware capabilities such as SR-IOV which
                                                                      splits one NVMe device into partitions and one NIC into
Table 1 summarizes OS design features and compares Li-
                                                                      logically separate NICs with dedicated Tx/Rx buffers. We
brettOS against existing monolithic, library, and multiserver
                                                                      assume that the NIC firmware does not have fatal security
OSes. Since LibrettOS builds on NetBSD’s anykernels, it in-
                                                                      vulnerabilities – i.e., it prevents applications from adversely
herits a very large driver code base. Applications do not
                                                                      affecting one another. (VF isolation is already widely used
generally need to be modified as the user space environ-
                                                                      by guest VMs where similar security concerns exist.)
ment is mostly compatible with that of NetBSD, except for
scenarios described in Section 4.2. LibrettOS supports two               While some high-end NICs expose VFs using SR-IOV, the
paradigms simultaneously: applications with high perfor-              total number of VFs is typically limited, and only a small
mance requirements can directly access hardware resources             number of applications that need faster network access will
as with library OSes (or DPDK [84] and SPDK [77]), while              benefit from this hardware capability. When there are no
other applications can use servers that manage correspond-            available VFs for applications, or their direct exposure to ap-
ing resources as with multiserver OSes. Similar to MINIX              plications is undesirable due to additional security require-
3 [30], HelenOS [13], VirtuOS [62], and Snap [56], LibrettOS          ments, we use network servers. As in library OS mode,
supports failure recovery of servers, which is an important           Application 2’s data (see Figure 3) goes through libc sock-
feature of the multiserver OS paradigm. Additionally, Li-             ets and the entire network stack to generate L2 (data link
brettOS’s applications can dynamically switch between the             layer) network frames. However, unlike Application 1, the
library and multiserver OS modes at run-time.                         network frames are not passed to/from the NIC driver di-
   In Section 6, we provide a detailed discussion of the past         rectly, but rather relayed through a virtual interface (VIF).
and related work and contrast them with LibrettOS.                    The VIF creates an IPC channel with a network server where
                                                                      frames are additionally verified and rerouted to/from the
                                                                      NIC driver.
3     Design                                                             LibrettOS is mostly agnostic to the kind of hypervisor
                                                                      or microkernel used by rumprun. For easier prototyping
One of the defining features of LibrettOS is its ability to use       while preserving rich IPC capabilities and microkernel-like
the same software stacks and device drivers irrespective of           architecture, we used Xen, which runs rumprun instances in
whether an application accesses hardware devices directly             VMs. Unlike KVM, Xen has a microkernel-style design and
or via servers. In this section, we describe LibrettOS’s archi-       does not use any existing monolithic OS for the hypervisor
tecture, its network server, and dynamic switching capabil-           implementation. Since rumprun has already been ported to
ity.                                                                  the seL4 microkernel [14], and our core contributions are
                                                                      mostly orthogonal to the underlying hypervisor, the Libret-
                                                                      tOS design principles can be adopted in seL4 as well.
3.1    LibrettOS’s Architecture
                                                                         Scheduling in LibrettOS happens at both the hypervi-
Figure 3 presents LibrettOS’s architecture. In this illus-            sor and user application levels. The hypervisor schedules
tration, Application 1 runs in library OS mode, i.e., it ac-          per-VM virtual CPUs (VCPUs) as usual. Whereas the orig-
cesses hardware devices such as NVMe and NIC directly                 inal rumprun implemented a simple non-preemptive N : 1
by running corresponding drivers in its own address space.            uniprocessor scheduler, we had to replace it entirely to sup-

                                                                  5
port SMP – i.e., multiple VCPUs per instance. Each rumprun               Application   mbuf            Network          mbuf   Application
                                                                              1      exchange           Server        exchange      2
instance in LibrettOS runs its own M : N scheduler that ex-
ecutes M threads on top of N VCPUs.                                          libc            Rx        handler 1         Rx           libc

   LibrettOS does not have any specific restrictions on the                TCP/IP    Rx:threads        handler 2     Rx:threads     TCP/IP
type of hardware used for direct access; we tested and eval-                                 Tx                           Tx
                                                                             VIF                       NIC driver                     VIF
uated both networking and storage devices. We only im-
                                                                                     Tx:threads                      Tx:threads
plemented a network server to demonstrate indirect access.                 rumprun                      rumprun                     rumprun
                                                                                        VIRQ1                          VIRQ2
However, storage devices and file systems can also be shared
using an NFS server. The NFS rumprun instance runs on top                               Figure 4. Network server.
of an NVMe device initialized with the ext3 file system and
exports file system shares accessible through network. In
                                                                                            Network Sever             Application
the long term, an existing rump kernel-based approach [41]
for running file systems in the user space of monolithic OSes                        Packet forward service               libc

                                                                                     READ
can be adopted in the rumprun environment as well.                                           RO-Mapped Pages
                                                                                                                       TCP/UDP
   LibrettOS does not currently support anything outside of                                 Portmap Table
the POSIX/BSD APIs. Admittedly, POSIX has its downsides
– e.g., copying overheads which are undesirable for high-                                                     Xen
performance networking. Nonetheless, LibrettOS already                                      Portmap Table   Portbind Hypercall
eliminates the system call layer, which can open the way for
further optimizations by extending the API in the future.
                                                                                            Figure 5. Portmap table.

3.2    Network Server
                                                                      routing based on TCP/UDP ports. When applications re-
Due to flexibility of rump kernels, their components can
                                                                      quest socket bindings, the network server allocates corre-
be chosen case by case when building applications. Unlike
                                                                      sponding ports (both static and dynamic) by issuing a spe-
more traditional OSes, LibrettOS runs the TCP/IP stack di-
                                                                      cial portbind hypercall as shown in Figure 5 which updates
rectly in the application address space. As we discuss be-
                                                                      a per-server 64K port-VM (portmap) table in the hypervi-
low, this does not compromise security and reliability. In
                                                                      sor. We store the portmap in the hypervisor to avoid its
fact, compared to typical multiserver and monolithic OSes,
                                                                      loss when the network server is restarted due to failures.
where buggy network stacks affect all applications, Libret-
                                                                      While this mechanism is specific to the network server, the
tOS’s network stack issues are localized to one application.
                                                                      added code to the hypervisor is very small and is justified
   The network server IPC channel consists of Tx/Rx ring
                                                                      for omnipresent high-speed networking. Alternatively, the
buffers which relay L2 network frames in mbufs through a
                                                                      portmap table can be moved to a dedicated VM.
virtual network interface (VIF) as shown in Figure 4. We use
a lock-free ring buffer implementation from [61]. When re-               Since packet routing requires frequent reading of the
laying mbufs, the receiving side, if inactive, must be notified       portmap entries, we read-only map the portmap pages into
using a corresponding notification mechanism. For the Xen             the network server address space, so that the network server
hypervisor, this is achieved using a virtual interrupt (VIRQ).        can directly access these entries.
   Since a stream of packets is generally transmitted without            The network server does not trust applications for se-
delays, the number of such notifications is typically small.          curity and reliability reasons. We enforce extra metadata
The sending side detects that the receiving side is inactive          verification while manipulating ring buffers. Since applica-
by reading a corresponding shared (read-only) status vari-            tions may still transmit incorrect data (e.g., spoof MAC/IP
able denoted as threads, which indicates how many threads             source addresses), either automatic hardware spoofing de-
are processing requests. When this variable is zero, a VIRQ           tection must be enabled, or address verification must be en-
is sent to wake a worker thread up.                                   forced by verifying MAC/IP addresses in outbound frames.
   The network server side has a per-application handler              Similarly, ports need to be verified by the server.
which forwards mbufs to the NIC driver. Frames do not                    Although, unlike POSIX, invalid packets may be gener-
need to be translated, as applications and servers are built          ated by (malicious) applications at any point, they can only
from the same rump kernel source base.                                appear in local (per-application) ring buffers. When the net-
   An important advantage of this design is that an applica-          work server copies packets to the global medium, it verifies
tion can recover from any network server failures, as TCP             MAC, IP, and port metadata which must already be valid for
state is preserved inside the application. The network server         a given application. Races are impossible as prior users of
operates at the very bottom of the network stack and sim-             the same MAC, IP, and port must have already transmitted
ply routes inbound and outbound mbufs. However, rout-                 all packets to the global medium and closed their connec-
ing across different applications requires some knowledge             tions.
of the higher-level protocols. We support per-application                We also considered an alternative solution based on the

                                                                  6
netfront/netback model [4] used by Xen. The netback driver                as connections are closed. When launching an application
typically runs in an OS instance with a physical device                   that needs dynamic switching (Application 2), the network
(Driver Domain). The netfront driver is used by guest OSes                server allocates an IP address that will be used by this ap-
as a virtual NIC to establish network connectivity through                plication. Applications use virtual interfaces (VIF) which
the netback driver. Rumprun already implements a very                     are initially connected to the network server (i.e., similar to
simple netfront driver so that it can access a virtual NIC                Application 1). The network server adds the IP to the list of
when running on top of the Xen hypervisor. However,                       configured addresses for its IF (PF or VF).
it lacks the netback driver that would be required for the                    As I/O load increases, an application can be decoupled
server. This model is also fundamentally different from                   from the network server to use direct access for better
our approach: transmitted data must traverse additional OS                throughput and CPU utilization. The NIC interface used by
layers on the network server side and all traffic needs to                the network server deactivates application’s IP (which re-
go through network address translation (NAT) or bridging,                 mains unchanged throughout the lifetime of an application).
as each application (VM instance) gets its own IP address.                The application then configures an available NIC interface
While the Xen approach is suitable when running guest                     (typically, VF) with its dedicated IP. Changing IP on an in-
OSes, it introduces unnecessary layers of indirection and                 terface triggers ARP invalidation. A following ARP probe
overheads as we show in Section 5.3.                                      detects a new MAC address for the interface-assigned IP.
                                                                          These events are masked from applications, as they typi-
                                                                          cally use higher-level protocols such as TCP/IP. No pause-
3.3     Dynamic Mode Switch                                               and-drain is needed, and on-the-fly requests are simply ig-
LibrettOS’s use of the same code base for both direct and                 nored because TCP automatically resends packets without
server modes is crucial for implementing the dynamic mode                 ACKs.
switch. We implement this feature for networking, which is                    When running in this mode, all traffic from the NIC in-
more or less stateless. As NVMe’s emerging support for SR-                terface (direct access) is redirected through VIF, as all appli-
IOV becomes publicly available, our technique can also be                 cation socket bindings belong to this VIF. This special VIF
considered for storage devices. However, storage devices                  is very thin: mbufs are redirected to/from IF without copy-
are more stateful and additional challenges need to be over-              ing, i.e., avoiding extra overhead. When the load decreases,
come.                                                                     applications can return to the original state by performing
   Figure 6 shows LibrettOS’s dynamic switching mecha-                    the above steps in the reverse order. If applications always
nism for the network server. LibrettOS is the only OS that                use direct access (Application 3), VIFs are not needed, i.e.,
supports this mechanism. Moreover, applications do not                    all traffic goes directly through IF.
need to be modified to benefit from this feature. As we fur-                  Dynamic switching helps conserve SR-IOV resources, i.e.,
ther discuss in Section 5, dynamic switching does not incur               it enables more effective use of available hardware resources
large overhead, and switching latency is small (56-69ms).                 across different applications. Policies for switching (e.g., for
                                                                          managing resources, security, etc.) are beyond the scope of
   Application      Network             Application
                                                                          this paper, as our focus is on the switching mechanism. In
                                                      Application
        1            Server                  2             3              our evaluation, users manually trigger dynamic switching.
         IP 1         IP 1               MAC/IP
                                           IP 2 2         IP 3

        VIF 1         IP 2
                              Dynamic
                                           VIF
                                            VF2            IF             4     Implementation
                               switch
                      IF                    IF        Direct access
      Uses server                                          only           While NetBSD code and its rump kernel layer are SMP-
                                                                          aware (except rump kernel bugs that we fixed), none of the
          Figure 6. Dynamic switching mechanism.                          existing rumprun versions (Xen PV, KVM, or seL4) support
                                                                          SMP.
   The biggest challenge for dynamic switching is that ap-
plication migration must be completely transparent. When
                                                                          4.1    Effort
using direct access (VF), we need a dedicated IP address for
each NIC interface. On the other hand, our network server                 We redesigned low-level parts of rumprun (SMP BIOS
shares the same IP address across all applications. A com-                initialization, CPU initialization, interrupt handling) as
plicating factor is that once connections are established, IP             well as higher-level components such as the CPU sched-
addresses used by applications cannot be changed easily.                  uler. Our implementation avoids coarse-grained locks
   To allow dynamic switching, we leverage NetBSD’s sup-                  on performance-critical paths. Additionally, the original
port for multiple IPs per an interface (IF). Unlike hard-                 rumprun was only intended for paravirtualized Xen. Thus,
ware Rx/Tx queues, IPs are merely logical resources; they                 we added support for SMP-safe, Xen’s HVM-mode event
are practically unlimited when using NAT. A pool of avail-                channels (VIRQs) and grant tables (shared memory pages)
able IPs is given to a server, and addresses are recycled                 when implementing our network server. We implemented

                                                                      7
Component           Description                   LOC                Processor         2 x Intel Xeon Silver 4114, 2.20GHz
                                                                       Number of cores   10 per processor, per NUMA node
  SMP support     SMP BIOS init/trampoline,          2092
                                                                       HyperThreading    OFF (2 per core)
                  M:N scheduler, interrupts
                                                                       TurboBoost        OFF
  HVM support     Xen grant tables and VIRQ          1240
                                                                       L1/L2 cache       64 KB / 1024 KB per core
  PV clock        SMP-safe Xen/KVM clock              148
                                                                       L3 cache          14080 KB
  Network server  mbuf routing,                      1869
                                                                       Main Memory       96 GB
                  dynamic switching
                                                                       Network           Intel x520-2 10GbE (82599ES)
  Xen             Network server registrar            516
                                                                       Storage           Intel DC P3700 NVMe 400 GB
  NetBSD, network Network server forwarder             51
  NetBSD, kqueue EVFILT_USER support                  303
  NetBSD, rump    Race conditions in SMP               45
                                                                                  Table 3. Experimental setup.
  NetBSD, glue    New drivers (NVMe, ixgbe)           234
  NFS, rpcbind,   Porting from NetBSD,                154
  and mountd      avoiding fork(2)                                 Moreover, typical applications that rely on fork [82] can be
  Total:                                            6652           configured to use the multi-threading mode instead.
                                                                      We did not implement POSIX’s IPCs such as pipe(2), but
                                                                   Xen-based mechanisms can be used with no restrictions.
            Table 2. Added or modified code.                          As fork(2) is not currently supported, we did not consider
                                                                   sharing socket file descriptors or listen(2)/accept(2) handoff
                                                                   across processes. This may create additional challenges in
the Xen HVM interrupt callback mechanism by binding it to          the future. We believe that the problem is tractable, but re-
the per-CPU interrupt vectors as it is implemented in Linux        quires modification of the network stack. Also, Rx/Tx ring
and FreeBSD. The HVM mode uses hardware-assisted virtu-            buffers will have to be created per connection rather than
alization and is more preferable nowadays for security rea-        per process to allow sharing per-socket/connection buffers
sons [53]. Moreover, it has better IOMMU support when us-          across processes while not violating process isolation re-
ing device PCI passthrough. We also fixed race conditions          quirements.
in the rump layer, and added NVMe and 10GbE ixgbe driver              As previously mentioned, we did not develop policies for
glue code which was previously unavailable for rumprun.            dynamic switching as our focus was on mechanism. Sec-
Additionally, we ported a patch that adds support for the          tion 5 uses a simple policy wherein a user triggers an event.
EVFILT_USER extension in kqueue(2), which is required
by some applications and supported by FreeBSD but is, for
some reason, still unsupported by NetBSD.                          5     Evaluation
   Since rumprun does not currently support preemption,
we designed a non-preemptive, lock-free M:N scheduler              We evaluate LibrettOS using a set of micro- and macro-
with global and per-VCPU queues. A lock-free scheduler             benchmarks and compare it against NetBSD and Linux.
has an advantage that, it is not adversely affected by pre-        Both of these baselines are relevant since LibrettOS is di-
emption which still takes place in the hypervisor. (When the       rectly based on the NetBSD code base, and Linux is a
hypervisor de-schedules a VCPU holding a lock, other VC-           popular server OS. LibrettOS typically has better perfor-
PUs cannot get the same lock and make further progress.)           mance, especially when using direct mode. Occasionally,
Lock-free data structures for OS schedulers were already ad-       Linux demonstrates better performance since it currently
vocated in [45].                                                   has more optimizations than NetBSD in its drivers and net-
   In Table 2, we present the effort that was required to          work stack, especially for 10GbE+, as we further discuss in
implement our current LibrettOS prototype. Most changes            Section 5.1.
pertain to rumprun and the network server implementation,             In the evaluation, we only consider standard POSIX com-
whereas drivers and applications did not require substantial       patible applications, as compatibility is one of our impor-
changes.                                                           tant design goals. Since kernel-bypass libraries and typical
                                                                   library OSes [7, 77, 84] are not compatible with POSIX, they
                                                                   cannot be directly compared. Moreover, kernel-bypass li-
4.2   Limitations
                                                                   braries require special TCP stacks [40] and high engineer-
In our current prototype, applications cannot create pro-          ing effort to adopt existing applications [89]. Likewise, com-
cesses through fork(2), as its implementation would require        parison against existing multiserver designs [30, 56] is chal-
complex inter-VM interactions which Xen is not designed            lenging due to very limited application and device driver
for. We consider such an implementation to be realistic            support.
for microkernels such as seL4. Each application, however,             Table 3 shows our experimental setup. We run Xen
can already support virtually unlimited number of threads.         4.10.1, and Ubuntu 17.10 with Linux 4.13 as Xen’s Dom0

                                                               8
(for system initialization only). We use the same version of                      5.2                         Sysbench/CPU
Linux as our baseline. Finally, we use NetBSD 8.0 with the
NET_MPSAFE feature5 enabled for better network scalabil-                          To verify our new scheduler and SMP support, we ran the
ity [58]. LibrettOS uses the same NetBSD code base, also                          Sysbench/CPU [80] multi-threaded test which measures the
with NET_MPSAFE enabled. We set the maximum trans-                                elapsed time for finding prime numbers. Overall, LibrettOS,
mission unit (MTU) to 9000 in all experiments to follow gen-                      Linux, and NetBSD show roughly similar performance.
eral recommendations [51] for the optimal 10GbE perfor-
mance. We measure each data point 10 times and present                            5.3                         NetPIPE
the average; the relative standard deviation is mostly less
than 1%.
   LibrettOS not only outperforms NetBSD, but also outper-                                                     8                      Linux

                                                                                        Throughput (Gbps)
forms Linux in certain tests. This is despite the fact that                                                    7            LibrettOS-Direct
                                                                                                               6           LibrettOS-Server
NetBSD can be slower than Linux. LibrettOS’s advantage                                                                              NetBSD
comes from its superior multiserver-library OS design and                                                      5               Rumprun-PV
                                                                                                               4
better resource management (compared to NetBSD).
                                                                                                               3
                                                                                                               2
                                                                                                               1
5.1     NetBSD and Linux performance                                                                           0
                                                                                                                        0.25     1     4    16     64          256
Since neither the original rumprun nor other unikernels                                                                           Message Size (KB)
support 10GbE+ or NVMe physical devices, one of our goals                                                                         (a) Throughput
is to show how this cutting-edge hardware works in Libret-                                                     100
tOS.
                                                                                        CPU utilization (%)        80
   Although NetBSD, Linux, and rumprun’s 1GbE are
known to be on par with each other [14], 10GbE+ network-                                                           60
ing puts more pressure on the system, and we found occa-
                                                                                                                   40
sional performance gaps between NetBSD and Linux. How-
ever, they are mostly due to specific technical limitations,                                                       20
most of which are already being addressed by the NetBSD
                                                                                                                   0
community.
                                                                                                                          0.25     1    4    16    64          256
   One contributing factor is the network buffer size.                                                                             Message Size (KB)
NetBSD uses mbuf chains; each mbuf is currently limited
                                                                                                                         (b) CPU utilization (1 CPU is 100%)
to 512 bytes in x86-64. In contrast, Linux has larger (frame-
sized) sk_buff primitives. Since 512-byte units are too small
for 10GbE+ jumbo frames, we increased the mbuf size to                                                                  Figure 7. NetPIPE (log2 scale).
4K; this improved performance by ≈10% on macrobench-
marks. mbuf cannot be increased beyond the 4K page size                              To measure overall network performance, we use Net-
easily without bigger changes to the network stack, but this                      PIPE, a popular ping pong benchmark. It exchanges a fixed
is likely to be addressed by the NetBSD community sooner                          size message between servers to measure the latency and
or later.                                                                         bandwidth of a single flow. We evaluate Linux, LibrettOS
   Another contributing factor is the use of global SMP                           (direct access mode), LibrettOS (using our network server),
locks by NetBSD. Since NET_MPSAFE was introduced very                             NetBSD, and the original rumprun with the netfront driver.
recently, and the effort to remove the remaining coarse-                          On the client machine, Linux is running for all the cases. We
grained locks is still ongoing, Linux’s network stack and                         configure NetPIPE to send 10,000 messages from 64 bytes
network drivers may still have superior performance in                            to 512KB. Figure 7(a) shows the throughput for different
some tests which use high-end 10GbE+ networking.                                  message sizes. LibrettOS (direct access) achieves 7.4Gbps
   We also found that the adaptive interrupt moderation                           with 256KB. It also has a latency of 282µs for 256KB mes-
option in NetBSD’s ixgbe driver adversely affected perfor-                        sages. LibrettOS (network server) achieves 6.4Gbps with
mance in some tests such as NetPIPE (e.g., NetBSD was only                        256KB, and the latency of 256KB messages is 326µs. Linux
able to reach half of the throughput). Thus, we disabled it.                      achieves 6.5Gbps with 256KB messages and the latency of
                                                                                  318.9µs. NetBSD reaches 6.4Gbps in this test with the la-
                                                                                  tency of 326µs. The original rumprun with the netfront
   5 NET_MPSAFE, recently introduced in NetBSD, reduces global SMP
                                                                                  driver reaches only 1.0Gbps with 256KB messages and has
locks in the network stack. By and large, the network stack is the only
major place where global locks still remain. The ongoing effort will elimi-       a large latency of 1,983µs.
nate them.                                                                           Overall, all systems except the original rumprun have

                                                                              9
comparable performance. With smaller messages (<                                          Figure 9 shows the throughput with different concurrency
256KB), LibrettOS-Direct can be a bit faster than Linux or                                levels and file sizes. LibrettOS (direct access and server)
NetBSD. We attribute this improvement to the fact that                                    is always faster than NetBSD. Except very large blocks
for smaller messages, the cost of system calls is non-                                    (128KB), LibrettOS-Direct also outperforms Linux. Partic-
negligible [76]. In LibrettOS, all system calls are replaced                              ularly, for 8KB/50, LibrettOS (direct access) is 66% faster
with regular function calls. Even LibrettOS-Server avoids                                 than NetBSD and 27% faster than Linux. Even LibrettOS-
IPC costs, as the server and application VMs typically run                                Server outperforms Linux in many cases due to its opti-
on different CPU cores.                                                                   mized, fast IPC mechanism. We note that LibrettOS is able
                                                                                          to outperform Linux even though the original NetBSD is
                        LibrettOS                  Linux              NetBSD
                                                                                          slower than Linux in this test. We generally attribute this
                    5                                                                     improvement to the fact that LibrettOS removes the system
Throughput (GB/s)

                    4                                                                     call layer. Furthermore, compared to NetBSD, LibrettOS im-
                    3                                                                     proves scheduling and resource utilization. For very large
                    2                                                                     blocks (128KB), NetBSD’s network stack current limitations
                    1
                                                                                          outweigh other gains.
                                                                                             We also evaluated dynamic switching when running Ng-
                    0
                             4      8      16      32     64    128    256     512        inx. In Figure 10, LibrettOS-Hybrid represents a test with
                                                Block Size (KB)                           runtime switching from the server to direct mode and back
                                                                                          to the server mode (the concurrency level is 20). Overall,
                                        Figure 8. NFS server.                             Nginx spends 50% of time in the server mode and 50% in the
                                                                                          direct mode. LibrettOS-Hybrid’s average throughput shows
   LibrettOS-Direct is more efficient not just on throughput                              benefits of using this mode compared to LibrettOS-Server,
but, more importantly, on CPU utilization (Figure 7(b)). The                              Linux, and NetBSD. The relative gain (Figure 10) over Linux,
saved CPU resources can be used in myriad ways – e.g.,                                    let alone NetBSD, is quite significant in this test. We also
running more applications, obtaining higher overall perfor-                               measured the cost of dynamic switching to be 56ms from
mance. For LibrettOS-Server, there is a 1.8x gap in overall                               the server to direct mode, and 69ms in the opposite direc-
CPU consumption due to an extra layer where packets are                                   tion.
transferred from an application to the server. The same ar-
gument applies to other experiments in this section – e.g.,                               5.6   Memcached
dynamic switching.
                                                                                          We also tested LibrettOS’s network performance with mem-
                                                                                          cached [57], a well-known distributed memory caching sys-
5.4                     NFS Server                                                        tem, which caches objects in RAM to speed up database ac-
To evaluate NVMe storage, we run an NFS server mac-                                       cesses. We use memcached 1.4.33 on the server side. On the
robenchmark by using Sysbench/FileIO [80] from the client.                                client side, we run memtier_benchmark 1.2.15 [71] which
We measure mixed file I/O composed of page cache and stor-                                acts as a load generator. The benchmark performs SET and
age I/O because users benefit from the page cache in gen-                                 GET operations using the memcache_binary protocol. We
eral. We mount an NVMe partition initialized with the ext3                                use the default ratio of operations 1:10 to ensure a typi-
file system and export it through the NFS server. For Li-                                 cal, read-dominated workload. In Figure 11, we present the
brettOS, we use direct NIC and NVMe access. We mount                                      number of operations per second for a small and large block
the provided NFS share on the client side and run the Sys-                                sizes and using a different number of threads (1-20). Each
bench test with different block sizes (single-threaded). Fig-                             thread runs 10 clients, and each client performs 100,000 op-
ure 8 shows that for all block sizes, LibrettOS outperforms                               erations.
both NetBSD and Linux. LibrettOS is 9% consistently faster                                   LibrettOS-Direct and Linux show similar overall perfor-
than both of them. This is because, LibrettOS avoids the                                  mance. LibrettOS-Server reveals a runtime overhead since it
system call layer which is known to cause a non-negligible                                has to run a server in a separate VM. Finally, NetBSD shows
performance overhead [76].                                                                the worst performance as in the previous test.

5.5                     Nginx HTTP Server                                                 5.7   Redis
To run a network-bound macrobenchmark, we choose Ng-                                      Finally, we ran Redis, an in-memory key-value store used
inx 1.8.0 [60], a popular web server. We use the Apache                                   by many cloud applications [70]. We use Redis-server 4.0.9
Benchmark [81] to generate server requests from the client                                and measure the throughput of 1,000,000 SET/GET opera-
side. The benchmark is set to send 10,000 requests with var-                              tions. In the pipeline mode, the Redis benchmark sends re-
ious levels of concurrency on the client side. (In this bench-                            quests without waiting for responses which improves per-
mark, concurrency is the number of concurrent requests.)                                  formance. We set the pipeline size to 1000. We measure

                                                                                     10
the performance of Redis for various number of concurrent                                                              4KB                                                   8KB
connections (1-20). We show results for the SET/GET oper-                                           90                                                     180
ation (128 bytes); trends for other sizes are similar. Libret-                                      80                                                     160

                                                                       Throughput (MB/s)
tOS’s GET throughput (Figure 12) is close to Linux, but the                                         70                                                     140
                                                                                                    60                                                     120
SET throughput is smaller. NetBSD reveals very poor per-                                            50                                                     100
formance in this test, much slower than that of LibrettOS                                           40                                                      80
and Linux. We suspect that the current Redis implementa-                                            30                                                      60
                                                                                                              LibrettOS-Direct
                                                                                                    20       LibrettOS-Server                               40
tion is suboptimal for certain BSD systems, which results in                                        10                  Linux                               20
                                                                                                                      NetBSD
worse performance that we did not observe in other tests.                                            0                                                       0
                                                                                                             0 10 20 30 40 50 60                                     0 10 20 30 40 50 60

5.8    Failure Recovery and Software Up-                                                                               16KB                                                  32KB
                                                                                                    300                                                    500
       grades
                                                                                                    250

                                                                       Throughput (MB/s)
                                                                                                                                                           400
LibrettOS can recover from any faults occurring in its                                              200
servers, including memory access violations, deadlocks, and                                                                                                300
                                                                                                    150
interrupt handling routine failures. Our model assumes that                                                                                                200
                                                                                                    100
failures are localized to one server, i.e., do not indirectly
                                                                                                     50                                                    100
propagate elsewhere. Applications that use our network
server do not necessarily fail: the portmap state is kept in                                             0                                                  0
                                                                                                              0 10 20 30 40 50 60                                    0 10 20 30 40 50 60
the hypervisor, and if recovery happens quickly, TCP can
still simply resend packets without timing out (applications                                                           64KB                                                  128KB
keep their TCP states). Additionally, when upgrading the                                            700                                                    1000
                                                                       Throughput (MB/s)
server, an artificial fault can be injected so that an applica-                                     600
                                                                                                                                                            800
tion can use a newer version of the server without restarting                                       500
                                                                                                    400                                                     600
the application itself.
                                                                                                    300                                                     400
   To demonstrate LibrettOS’s failure recovery (or transpar-                                        200
ent server upgrades), we designed two experiments. In the                                                                                                   200
                                                                                                    100
first experiment, an application has two jobs: one uses net-                                          0                                                          0
working (Nginx) and the other one accesses a file system.                                                     0 10 20 30 40 50 60                                     0 10 20 30 40 50 60
When LibrettOS’s network server fails, the Nginx server is                                                        Concurrency                                             Concurrency

unable to talk to the NIC hardware, and we report outage
                                                                                                                     Figure 9. Nginx HTTP server.
that is observable on the client side. This failure, nonethe-
less, does not affect the file system job. In Figure 13(a), we
                                                                                                     LibrettOS-Direct                                        LibrettOS-Hybrid
show the corresponding network and file system transfer                                             LibrettOS-Server                                                    Linux
                                                                       Absolute Throughput (MB/s)

speeds. After rebooting the network server, Nginx can again                                                  NetBSD
                                                                                                                                    Gain (% of NetBSD’s)

                                                                                                                                                            50
reach the network, and clients continue their data transfers.                                       400
                                                                                                    350                                                     40
In the other test, Figure 13(b), we run two applications. Ng-                                       300
                                                                                                    250                                                     30
inx uses direct access while Redis uses our network server.                                         200
                                                                                                    150                                                     20
We show that a failure of Nginx does not affect Redis. Then                                         100                                                     10
we also show that a network server failure does not impact                                           50
                                                                                                      0                                                      0
Nginx.                                                                                                             8    16 32                                            8     16 32
                                                                                                                  File Size (KB)                                         File Size (KB)

6     Related Work                                                                                               Figure 10. Dynamic switch (Nginx).

LibrettOS’s design is at the intersection of three OS research
topics: system component isolation for security, fault-tole-           and reliability guarantees [22, 31, 46]. With microkernels,
rance, and transparent software upgrades (multiserver OS);             only essential functionalities (scheduling, memory manage-
application-specialized OS components for performance,                 ment, IPC) are implemented within the kernel. L4 [52]
security, and software bloat reduction (library OS); and ker-          is a family of microkernels, which is known to be used
nel bypassing with direct hardware access to applications              for various purposes. A notable member of this family
for performance.                                                       is seL4 [46], a formally verified microkernel. Multiserver
   Isolation of system software components in separate ad-             OSes such as MINIX 3 [30], GNU Hurd [8], Mach-US [79],
dress spaces is the key principle of the microkernel model [1,         and SawMill [20] are a specialization of microkernels where
13, 27, 30, 33, 36, 52]. Compared to the classical mono-               OS components (e.g., network and storage stacks, device
lithic OS design, microkernels provide stronger security               drivers) run in separate user processes known as “servers.”

                                                                  11
LibrettOS-Direct                                                               Linux                      network stacks such as mTCP [40], MegaPipe [25], and
                                 LibrettOS-Server                                                              NetBSD
                                                  32 bytes                                                           8KB
                                                                                                                                            OpenOnload [65] have been designed to bypass the OS and
                                                                                                                                            avoid various related overheads. While it is not their pri-
                              250000
                                                                                           200000                                           mary objective, such user-space network stacks do provide
Operations / s

                              200000
                                                                                           150000                                           a moderate degree of reliability: since they do not reside in
                              150000
                                                                                           100000                                           the kernel, a fault in the network stack will not affect the
                              100000
                                                                                                   50000
                                                                                                                                            applications that are not using it.
                               50000
                                                                                                                                               In its multiserver mode, LibrettOS obtains the security
                                        0                                                                 0
                                        0    5 10 15 20                                                  0     5   10 15 20                 and reliability benefits of microkernels by decoupling the
                                    Threads (each runs 10 clients)                                    Threads (each runs 10 clients)        application from system components such as drivers that
                                                                                                                                            are particularly prone to faults [23, 32] and vulnerabili-
                              Figure 11. Memcached (distributed object system).                                                             ties [9].
                                                                                                                                               IX [7] and Arrakis [64] bypass traditional OS layers to im-
                                  LibrettOS-Direct                                                               Linux                      prove network performance compared to commodity OSes.
                                 LibrettOS-Server                                                              NetBSD
                                             128 bytes (Set)                                                  128 bytes (Get)
                                                                                                                                            IX implements its own custom API while using the DPDK
                                                                                                    1.6                                     library [84] to access NICs. Arrakis supports POSIX but
Million transact / s

                                                                    Million transact / s

                                1                                                                   1.4                                     builds on top of the Barrelfish OS [5], for which device
                              0.8                                                                   1.2
                                                                                                      1                                     driver support is limited. LibrettOS gains similar perfor-
                              0.6                                                                   0.8                                     mance benefits when using direct hardware access (library
                              0.4                                                                   0.6
                                                                                                    0.4                                     OS) mode, while reducing code development and mainte-
                              0.2
                                                                                                    0.2                                     nance costs as it reuses a very large base of drivers and soft-
                                0                                                                     0
                                    0         5   10    15     20                                         0     5   10    15    20          ware from a popular OS, NetBSD. Additionally, in contrast
                                              Concurrency                                                       Concurrency                 to LibrettOS, these works do not consider recoverability of
                                                                                                                                            OS servers.
                                              Figure 12. Redis key-value store.                                                                Certain aspects of the multiserver design can be em-
                                                                                                                                            ployed using existing monolithic OSes. VirtuOS [62] is a
                                                                                                                                            fault-tolerant multiserver design based on Linux that pro-
          Throughput (MB/s)

                                                                               Throughput (MB/s)

                               1200                                                                  1200
                                                  NIC                                                         Nginx (Direct)
                               1000                FS
                                                                                                     1000     Redis (Server)                vides strong isolation of the kernel components by running
                                800                                                                   800
                                600                                                                   600                                   them in virtualized service domains on top of Xen. Snap [56]
                                400                                                                   400                                   implements a network server in Linux to improve perfor-
                                200                                                                   200
                                  0                                                                     0                                   mance and simplify system upgrades. However, Snap uses
                                            0 5 10 15 20 25 30                                                0 10 20 30 40 50              its private (non-TCP) protocol which cannot be easily inte-
                                               Time (sec)                                                        Time (sec)
                                                                                                                                            grated into existent applications. Moreover, Snap is limited
                       (a) Single application (Nginx)                                              (b) Multiple applications                to networking and requires re-implementing all network
                                                                                                                                            device drivers.
                Figure 13. Failure recovery and transparent upgrades.                                                                          On-demand virtualization [47] implements full OS migra-
                                                                                                                                            tion, but does not specifically target SR-IOV devices. In con-
                                                                                                                                            trast, LibrettOS allows dynamic device migrations.
   Decomposition and isolation have proved to be benefi-                                                                                       To optimize I/O in VMs, researchers have proposed the
cial for various layers of system software in virtualized en-                                                                               sidecore approach [19, 48] which avoids VM exits and of-
vironments: hypervisor [75], management toolstack [11],                                                                                     floads the I/O work to sidecores. Since this approach is
and guest kernel [62]. Security and reliability are partic-                                                                                 wasteful when I/O activity reduces, Kuperman et al. [49]
ularly important in virtualized environments due to their                                                                                   proposed to consolidate sidecores from different machines
multi-tenancy characteristics and due to the increased re-                                                                                  onto a single server.
liance on cloud computing for today’s workloads. In the                                                                                        Exokernel [15] was among the first to propose compiling
desktop domain, Qubes OS [73] leverages Xen to run ap-                                                                                      OS components as libraries and link them to applications.
plications in separate VMs and provides strong isolation for                                                                                Nemesis [72] implemented a library OS with an extremely
local desktop environments. LibrettOS also uses Xen to play                                                                                 lightweight kernel. Drawbridge [66], Graphene [85], and
the role of the microkernel. However, the LibrettOS design                                                                                  Graphene-SGX [86] are more recent works that leverage
is independent of the microkernel used and can be applied                                                                                   the security benefits of library OSes due to the increased
as-is to other microkernels supporting virtualization, with                                                                                 isolation resulting from the lesser degree of interaction be-
potentially smaller code bases [46, 78, 87].                                                                                                tween applications and the privileged layer. Bascule [6]
   The network stack is a crucial OS service, and its reliabil-                                                                             demonstrated OS-independent extensions for library OSes.
ity has been the subject of multiple studies based on micro-                                                                                EbbRT [74] proposed a framework for building per-app-
kernels/multiserver OSes using state replication [12], par-                                                                                 lication library OSes for performance.
titioning [35], and checkpointing [21]. Several user-space                                                                                     Library OSes specialized for the cloud – unikernels [10,

                                                                                                                                       12
You can also read